Object recognition by a self-organizing neural network which grows adaptively

We describe a new type of neural network for object recognition which we call a Cresceptron. The term “Cresceptron” was coined from Latin cresco (grow) and perceplio (perception). The primary objective of the Cresceptron framework is to automatically handle manually intractable tasks: such as constructing a network that can recognize many objects from real world images. The Cresceptron uses a hierarchical structure, and the network adaptivcly and incrementally grows through learning. For recognition, the network is made largely translationally invariant by using the same neuron at all the positions of each neural plane. Scale invariance is achieved through a multi-resolution representation with the framework of visual attention. Limited oricntational invariance is obtained by variation tolerance. Complete oricnlational invariance is not sought here since die recognition should report also the orientation. It is interesting to note that psychophysical studies have demonstrated that the human vision system does not have perfect invariance in cither translation, scale, or orientation.